Visitor identification and multi-dimensional user portrait routing allocation method based on private domain traffic

Abstract

The present invention provides a visitor identification and multi-dimensional user portrait routing allocation method based on private domain traffic, which relates to the field of data analysis technology, including collecting multi-source data information of visitors from a private domain traffic pool, constructing a multi-dimensional feature vector, generating visitor identity fingerprint features, and establishing a visitor identity recognition model based on the visitor identity fingerprint features; constructing a multi-level label system based on the results of the visitor identity recognition model; using a transfer learning algorithm to extract features from the visitor's historical behavior trajectory data and historical content interaction data, classifying the extracted features based on the multi-level label system to generate user portrait features, dynamically updating the user portrait features through incremental learning, constructing a multi-dimensional user portrait model, and combining the real-time load status of the service node, using a routing allocation strategy to allocate different visitor requests to corresponding service nodes.

Description

Technical Field

[0001] The present invention relates to data analysis technology, and in particular to a visitor identification and multi-dimensional user portrait routing allocation method based on private domain traffic. Background Art

[0002] Traditional visitor request processing methods mainly use polling or random allocation load balancing strategies, which fail to fully utilize visitor feature information and user portrait data, resulting in inefficient service resource allocation and poor user experience.

[0003] In the existing technology, visitor identification usually relies on a single device ID or cookie information, which makes it difficult to deal with complex scenarios such as cross-device access and cookie clearing. At the same time, the process of building user portraits has problems such as high data dimensions, sparse features, and poor real-time performance, which affects the accuracy of user feature expression and calculation efficiency.

[0004] In terms of service routing allocation, the following technical problems exist: First, routing decisions fail to effectively integrate user profile features and service node load status, resulting in unbalanced resource allocation; second, static routing strategies cannot adapt to dynamic changes in business loads, affecting system scalability; third, there is a lack of real-time monitoring and feedback mechanisms for service quality and user experience, making it difficult to ensure the continued stability of services. In addition, existing routing allocation solutions often separate load balancing from feature matching and lack an effective fusion mechanism. In high-concurrency scenarios, problems such as severe uneven load on service nodes or improper handling of user requests may occur, affecting the overall performance and service quality of the system. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for visitor identification and multi-dimensional user portrait routing allocation based on private domain traffic, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides visitor identification and multi-dimensional user portrait routing allocation methods based on private domain traffic, including:

[0008] Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features;

[0009] Based on the results of the visitor identification model, a multi-level label system is constructed, and the multi-level label system includes basic attribute labels, behavior feature labels, consumption ability labels, interest preference labels, social relationship labels, and time period labels; a transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features, and the user portrait features are dynamically updated through incremental learning to construct a multi-dimensional user portrait model;

[0010] According to the user portrait characteristics of the multi-dimensional user portrait model and the real-time load status of the service nodes, a routing allocation strategy is adopted to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service nodes is monitored. When load imbalance is detected, the routing allocation strategy is dynamically adjusted.

[0011] In an optional embodiment,

[0012] Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features, including the following steps:

[0013] Preprocessing the multi-source data information, wherein the time holes in the behavior trajectory data are filled by an interpolation method based on temporal correlation, the abnormal parameters in the device feature data are detected and processed by the interquartile range method, the numerical features are normalized by minimum and maximum, and the text data are segmented and vectorized;

[0014] Constructing a multidimensional feature vector based on the preprocessed multi-source data information, the multidimensional feature vector includes time series features, spatial features and behavioral features, wherein the time series features of behavior frequency and duration are extracted from the behavioral trajectory data, the spatial features of activity areas and device migration trajectories are constructed using the device feature data and geographic location data, and the behavioral features of content preferences and interaction patterns are extracted in combination with the social relationship data and content interaction data;

[0015] The multi-dimensional feature vector is subjected to feature fusion by using a multi-branch deep neural network model, wherein the multi-branch deep neural network model includes a temporal feature branch, a spatial feature branch and a behavioral feature branch, wherein the temporal feature branch uses a bidirectional long short-term memory network to process temporal features, the spatial feature branch uses a convolutional neural network to process spatial features, and the behavioral feature branch uses an attention mechanism to process behavioral features, and the temporal features, spatial features and behavioral features are fused by a weight fusion mechanism;

[0016] An improved density-based spatial clustering algorithm is used to perform cluster analysis on the fused features. The improved density-based spatial clustering algorithm includes: using a kernel density estimation method to adaptively calculate a neighborhood radius parameter, dynamically adjusting a minimum sampling point threshold based on a time window sliding average method, introducing a timestamp-based neighborhood constraint, and generating a visitor identity fingerprint feature that includes visitor behavior patterns, activity patterns, and social characteristics;

[0017] A visitor identification model is established based on the visitor identity fingerprint feature, and the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier.

[0018] In an optional embodiment,

[0019] A visitor identification model is established based on the visitor identity fingerprint feature, wherein the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier, and the steps include:

[0020] A two-level ensemble learning framework is constructed, wherein the first level of the two-level ensemble learning framework includes a random forest classifier and an extreme gradient boosting classifier, wherein the random forest classifier uses a parameter optimization method based on an outsourcing error rate to determine the number of trees and depth parameters, and the extreme gradient boosting classifier uses a Bayesian optimization algorithm to determine the learning rate and subsampling ratio parameters; based on the prediction outputs of the two classifiers, a weighted voting fusion module is constructed in the second level of the two-level ensemble learning framework, wherein the weighted voting fusion module calculates the prediction result weights of the two classifiers by minimizing the logarithmic loss function, combines the prediction results of the random forest classifier and the extreme gradient boosting classifier according to the optimal weights, and introduces a regularization term in the logarithmic loss function to balance the weight distribution of the two classifiers;

[0021] A time-aware cross-validation strategy is used to train the visitor identification model. The data set is divided into multiple continuous subsets in chronological order. A feature perturbation enhancement mechanism and a sample resampling mechanism are introduced during the training process. The feature perturbation enhancement mechanism adds random noise to numerical features, and the sample resampling mechanism balances the sample distribution by generating synthetic samples.

[0022] An online update mechanism is designed to realize the incremental learning of the visitor identity recognition model. A sample buffer pool based on a sliding time window is maintained. When the number of new samples reaches the preset condition, the incremental learning process is triggered. A progressive update strategy is used to update the parameters of some weak learners. The resampling parameters are dynamically adjusted based on the data distribution of the sample buffer pool, so that the visitor identity recognition model can adapt to the dynamic changes of visitor behavior patterns.

[0023] In an optional embodiment,

[0024] Based on the result of the visitor identification model, a multi-level label system is constructed, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption capacity labels, interest preference labels, social relationship labels, and time period labels. The steps include:

[0025] Collect the visitor's device information, IP address and access time, extract the device hardware parameters to form a device feature vector, parse the geographic location coordinates based on the IP address to form a location feature vector, divide the access time into time periods to form a time feature vector, and combine the device feature vector, location feature vector and time feature vector to construct a basic attribute label;

[0026] Applying a bidirectional long short-term memory network to the visitor's page browsing sequence to extract the time series pattern of the behavior trajectory data, calculating the behavior indicators of page dwell time, jump frequency and repeat visit rate, and constructing a behavior feature label based on the time series pattern and behavior indicators, wherein the behavior feature label includes the visitor's browsing type and target characteristics;

[0027] Based on the time feature vector in the basic attribute tag and the browsing type in the behavior feature tag, combined with the transaction amount and purchase frequency of the visitor, a clustering algorithm is used to divide the consumption capacity groups, construct a price sensitivity index, and form a consumption capacity tag;

[0028] Use convolutional neural networks to process visitors' content interaction data to obtain spatial features, extract the distribution of stay time in different content areas, calculate content preference vectors based on click, collection and sharing behaviors, and construct interest preference tags;

[0029] Based on the attention mechanism, the behavioral feature labels are processed to extract the interaction intensity between visitors, build a visitor social graph and calculate the centrality and influence characteristics, and apply the community discovery algorithm to identify the connected visitor groups to form social relationship labels;

[0030] Perform Fourier transform on the access time series to extract periodic features, calculate activity indicators based on the sliding time window, and construct time period labels by combining the time feature vector in the basic attribute label and the target feature in the behavior feature label.

[0031] In an optional embodiment,

[0032] The steps of extracting features from the visitor's historical behavior trajectory data and historical content interaction data by using a transfer learning algorithm, classifying the extracted features based on the multi-level label system to generate user portrait features, and dynamically updating the user portrait features by using an incremental learning method to construct a multi-dimensional user portrait model include:

[0033] A transfer learning algorithm is used to extract features of visitors' historical behavior trajectory data and historical content interaction data. The historical behavior trajectory data includes click sequence, browsing path and stay time data, and the historical content interaction data includes search terms, comment text and interaction record data. The behavior features and content features are processed separately through a two-stream feature extraction network, and the obtained behavior features and content features are fused through an attention mechanism to obtain a unified feature representation;

[0034] Building a feature mapping network based on a multi-level tag system, decomposing the unified feature representation according to the tag level, building a mapping relationship matrix between features and tags, calculating the weight distribution of features in each tag dimension, and generating structured user portrait features;

[0035] Collect new behavior trajectory data and content interaction data of visitors in real time, design feature importance evaluation indicators, and dynamically update the user portrait features through incremental learning;

[0036] Based on the dynamically updated user portrait features and the feature importance evaluation index, the weight coefficient of the feature mapping network is updated to construct a multi-dimensional user portrait model.

[0037] In an optional embodiment,

[0038] According to the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service node, a routing allocation strategy is used to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service node is monitored. When load imbalance is detected, the steps of dynamically adjusting the routing allocation strategy include:

[0039] A dimensionality reduction algorithm is used to reduce the dimension of user portrait features. Based on the reduced dimensionality user portrait features, a clustering algorithm is used to divide users into multiple clusters, and the mean vector of user portrait features in each cluster is calculated as the representative feature of this type of user;

[0040] Obtain user profile features of new visitors, calculate the matching degree between the user profile features and the representative features of each cluster, and introduce a load balancing factor, wherein the load balancing factor is calculated based on the current load of each service node and the maximum load allowed by the system;

[0041] Constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to a corresponding service node based on the routing decision score;

[0042] The real-time load status of the service node is monitored, and the load standard deviation is calculated. When the load standard deviation exceeds a preset threshold, the routing allocation strategy is dynamically updated by adjusting the weight coefficient of the matching degree and the load balancing factor.

[0043] In an optional embodiment,

[0044] The steps of constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to the corresponding service node based on the routing decision score include:

[0045] Obtain a user portrait feature vector and a representative feature vector, and calculate a feature dimension weight, wherein the feature dimension weight is obtained by an information gain method based on the difference between the user category information entropy and the conditional entropy; perform a weighted calculation on the user portrait feature vector and the representative feature vector based on the feature dimension weight to obtain a weighted cosine matching degree;

[0046] Obtaining the current load level of each service node, and normalizing the current load level using a nonlinear function to obtain a load balancing factor;

[0047] Constructing a load balancing loss function including a load variance term and a feature matching deviation term, and calculating a time-varying fusion weight using a gradient descent method based on the load balancing loss function;

[0048] The weighted cosine matching degree and the load balancing factor are weightedly combined by the time-varying fusion weight to obtain a routing decision score;

[0049] The routing decision score is converted into a routing probability distribution by using a soft maximization method, and a selection result of a service node is determined based on the routing probability distribution.

[0050] According to a second aspect of the embodiments of the present invention,

[0051] Provides a visitor identification and multi-dimensional user portrait routing distribution system based on private domain traffic, including:

[0052] The first unit is used to collect multi-source data information of visitors from a private domain traffic pool, wherein the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data, and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features, and behavior features, perform feature fusion and cluster analysis on the multi-dimensional feature vector using a multi-branch deep neural network model, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features;

[0053] The second unit is used to construct a multi-level label system based on the result of the visitor identity recognition model, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption ability labels, interest preference labels, social relationship labels, and time period labels; a transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features, and the user portrait features are dynamically updated through incremental learning to construct a multi-dimensional user portrait model;

[0054] The third unit is used to allocate different visitor requests to corresponding service nodes based on the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service nodes using a routing allocation strategy, and monitor the real-time load status of the service nodes at the same time. When load imbalance is detected, the routing allocation strategy is dynamically adjusted.

[0055] According to a third aspect of the embodiments of the present invention,

[0056] An electronic device is provided, comprising:

[0057] processor;

[0058] a memory for storing processor-executable instructions;

[0059] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0060] A fourth aspect of the embodiments of the present invention is:

[0061] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0062] The present invention realizes accurate identification of visitors through multi-source data fusion and deep learning; the multi-level label system and transfer learning improve the accuracy and real-time performance of user portraits; the intelligent routing allocation based on portrait features significantly improves the system performance and effectively enhances the operational effect and user experience of private domain traffic.

[0063] The two-level integrated learning framework of the present invention fully integrates the advantages of different classifiers and improves the generalization ability of the model; the time-series-aware cross-validation strategy ensures the time-series consistency of model evaluation; the feature perturbation and sample resampling mechanism enhance the robustness of the model; the online update mechanism realizes the continuous optimization of the model, enabling it to adapt to the changes in visitor behavior patterns in a timely manner. The overall solution shows significant technical advantages in the visitor identification scenario of private domain traffic, providing reliable technical support for precision marketing and personalized services.

[0064] The present invention achieves a comprehensive characterization of visitor characteristics through the construction of a multi-level label system; the feature extraction method based on deep learning improves the expressiveness of labels; the multi-dimensional label division meets the personalized needs of precision marketing; and the dynamic update mechanism ensures the timeliness of the label system. This solution provides systematic label support for the refined operation of private domain traffic, and can effectively guide the formulation of personalized recommendations and precision marketing strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 A flow chart of a method for visitor identification and multi-dimensional user portrait routing allocation based on private domain traffic according to an embodiment of the present invention;

[0066] Figure 2 This is a structural diagram of a visitor identification and multi-dimensional user portrait routing distribution system based on private domain traffic in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0068] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0069] Figure 1 Schematic diagram of the process of visitor identification and multi-dimensional user portrait routing allocation method based on private domain traffic according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0070] S1. Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features;

[0071] S2. Based on the results of the visitor identification model, a multi-level label system is constructed, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption ability labels, interest preference labels, social relationship labels, and time period labels; a transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features, and the user portrait features are dynamically updated through incremental learning to construct a multi-dimensional user portrait model;

[0072] S3. Based on the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service nodes, a routing allocation strategy is used to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service nodes is monitored. When load imbalance is detected, the routing allocation strategy is dynamically adjusted.

[0073] Exemplarily, multi-source data collection and preprocessing technology first collects visitor data in the private domain traffic pool in an all-round way. Behavior trajectory data includes user interaction information such as page visit sequence, dwell time, and click events; device feature data covers hardware and software features such as terminal type, operating system version, and browser identification; geographic location data records spatial information such as access IP, GPS coordinates, and movement trajectory; social relationship data includes social network features such as friend relationships, interaction frequency, and influence; content interaction data counts content browsing, collections, and comments. The collected raw data is cleaned and standardized, outliers and missing values ​​are removed, and the data format and dimension are unified.

[0074] The feature engineering phase constructs a multi-dimensional feature vector. The temporal feature extracts time series patterns such as visit frequency and dwell time through a sliding time window. A typical scenario is that a visitor visits the product details page 20 times in 30 days and the average dwell time is 3 minutes. The spatial feature constructs an activity area heat map based on geographic location information. For example, 90% of a visitor's visits come from Zhongguancun Street, Haidian District, Beijing. The behavioral feature extracts behavioral paths and interest preferences from clickstream data, such as the conversion link characteristics of browsing-collection-ordering.

[0075] The identity recognition model adopts a multi-branch deep neural network architecture. The network input layer receives a multi-dimensional feature vector and processes temporal features, spatial features, and behavioral features through parallel feature extraction branches. Each branch contains multiple convolutional layers and pooling layers to extract local and global representations of features. The feature fusion layer splices the outputs of each branch, interacts with features through a fully connected layer and an attention mechanism, and finally outputs a visitor identity fingerprint feature vector. Cluster analysis clusters similar fingerprint features based on similarity metrics to achieve accurate visitor identity recognition.

[0076] The construction of the multi-level tag system adopts a top-down hierarchical design. Basic attribute tags include demographic characteristics; behavioral characteristic tags describe access habits and usage scenarios; consumption capacity tags evaluate payment ability based on transaction data; interest preference tags analyze interest points through content interaction behavior; social relationship tags describe social network positions; and time period tags describe behavioral patterns in the time dimension.

[0077] The transfer learning mechanism is introduced in the user portrait modeling stage. The pre-trained model extracts the common feature representations from the visitor's historical behavior data, and then fine-tunes it in combination with the target domain data to improve the efficiency of feature extraction. The incremental learning method supports the real-time update of portrait features. When new behavior data appears, the model can dynamically adjust the feature weights to maintain the timeliness of the portrait.

[0078] The routing allocation strategy implements intelligent scheduling based on user profile characteristics and load balancing. First, classify visitors based on user profiles to determine service priorities and resource requirements. Monitor the CPU usage, memory usage, request queue and other load indicators of each service node in real time. When load imbalance is detected, dynamically adjust the routing weight to allocate new visitor requests to nodes with lower loads. At the same time, set a load threshold. When the node load exceeds the threshold, trigger an alarm and start load balancing.

[0079] The present invention realizes accurate identification of visitors through multi-source data fusion and deep learning; the multi-level label system and transfer learning improve the accuracy and real-time performance of user portraits; the intelligent routing allocation based on portrait features significantly improves the system performance and effectively enhances the operational effect and user experience of private domain traffic.

[0080] In an optional embodiment,

[0081] Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features, including the following steps:

[0082] Preprocessing the multi-source data information, wherein the time holes in the behavior trajectory data are filled by an interpolation method based on temporal correlation, the abnormal parameters in the device feature data are detected and processed by the interquartile range method, the numerical features are normalized by minimum and maximum, and the text data are segmented and vectorized;

[0083] Constructing a multidimensional feature vector based on the preprocessed multi-source data information, the multidimensional feature vector includes time series features, spatial features and behavioral features, wherein the time series features of behavior frequency and duration are extracted from the behavioral trajectory data, the spatial features of activity areas and device migration trajectories are constructed using the device feature data and geographic location data, and the behavioral features of content preferences and interaction patterns are extracted in combination with the social relationship data and content interaction data;

[0084] The multi-dimensional feature vector is subjected to feature fusion by using a multi-branch deep neural network model, wherein the multi-branch deep neural network model includes a temporal feature branch, a spatial feature branch and a behavioral feature branch, wherein the temporal feature branch uses a bidirectional long short-term memory network to process temporal features, the spatial feature branch uses a convolutional neural network to process spatial features, and the behavioral feature branch uses an attention mechanism to process behavioral features, and the temporal features, spatial features and behavioral features are fused by a weight fusion mechanism;

[0085] An improved density-based spatial clustering algorithm is used to perform cluster analysis on the fused features. The improved density-based spatial clustering algorithm includes: using a kernel density estimation method to adaptively calculate a neighborhood radius parameter, dynamically adjusting a minimum sampling point threshold based on a time window sliding average method, introducing a timestamp-based neighborhood constraint, and generating a visitor identity fingerprint feature that includes visitor behavior patterns, activity patterns, and social characteristics;

[0086] A visitor identification model is established based on the visitor identity fingerprint feature, and the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier.

[0087] Exemplarily, the multi-source data information preprocessing link first optimizes the data quality. For time holes in behavioral trajectory data, such as missing access records between 10:00 and 14:00 for a user, an interpolation method based on time series correlation is used to reasonably fill in the gaps by analyzing the behavioral characteristics of the previous and next time points to obtain a complete behavioral sequence. The interquartile range method is used to detect abnormal parameters in device feature data. Taking CPU usage as an example, if a value exceeds the upper quartile by plus or minus 1.5 times the interquartile range, it is judged as abnormal and replaced by the adjacent valid value. Numerical features such as device models and operating system versions are normalized to the minimum and maximum values, and the values ​​are mapped to the range of 0-1. Text data such as user comments are segmented using word segmentation technology and then converted into word vector representation.

[0088] Three types of features are extracted during the construction of multidimensional feature vectors. Temporal feature extraction uses a 24-hour time window to calculate the frequency and duration of visits. For example, a visitor visits an average of 5 times per hour between 9:00 and 18:00 on weekdays, with a single stay of 3 minutes. In the construction of spatial features, the user activity area is divided into 100x100 grids, and the access popularity of each grid is counted, while the migration trajectory of the device between different locations is recorded. Behavioral feature extraction focuses on the interaction pattern between users and content, such as the frequency distribution of clicks, favorites, forwarding and other operations, as well as the intensity of interaction with friends in social networks.

[0089] Feature fusion adopts a multi-branch deep neural network architecture. The temporal feature branch uses a bidirectional long short-term memory network with an input dimension of 24 (hours) x10 (number of features) and 256 hidden units, which can consider both forward and backward temporal dependencies. The spatial feature branch uses a 4-layer convolutional neural network with a convolution kernel size of 3x3 and the number of channels of 32, 64, 128, and 256, respectively, to extract local and global patterns of spatial features. The behavioral feature branch introduces a multi-head attention mechanism with the number of heads set to 8 and the dimension of each attention head being 64, highlighting the weight of important behavioral features. The outputs of the three branches are combined through a weight fusion layer, and the weight coefficients are obtained through back-propagation optimization.

[0090] Cluster analysis uses an improved density-based spatial clustering algorithm. The kernel density estimation method dynamically calculates the neighborhood radius and adaptively adjusts it in different density areas. Sparse areas take a larger radius such as 0.1, and dense areas take a smaller radius such as 0.02. The minimum sampling point threshold is based on a 30-minute time window to calculate the sliding average, which reflects the changes in data distribution in real time. The timestamp constraint requires that the time interval between clustered sample points does not exceed 24 hours to ensure the time correlation of the clustering results. The identity fingerprint features obtained through cluster analysis include users' typical visit time periods, activity areas, behavior patterns, and other features.

[0091] The visitor identification model uses an integrated learning framework, integrating two classifiers: random forest and extreme gradient boosting. The random forest uses 100 decision trees, with a maximum depth of 10 for each tree and a feature sampling ratio of 0.8. The extreme gradient boosting classifier sets the learning rate to 0.1, the maximum number of iterations to 500, and the sampling ratio of each iteration to 0.7. The prediction results of the two base classifiers are integrated through a voting mechanism to improve the generalization ability of the model.

[0092] Through multi-source data collection and preprocessing technology, the present invention realizes the all-round collection and quality optimization of visitor data, providing a high-quality data basis for subsequent analysis; the construction of multi-dimensional feature vectors makes full use of temporal features, spatial features and behavioral features, and comprehensively describes the behavior patterns and activity patterns of visitors; the multi-branch deep neural network model realizes the effective fusion of different types of features, significantly improving the feature expression ability; the improved density-based spatial clustering algorithm can adaptively process data with different density distributions, improving the clustering effect; the visitor identity recognition model based on ensemble learning has strong generalization and anti-interference capabilities.

[0093] In an optional embodiment,

[0094] A visitor identification model is established based on the visitor identity fingerprint feature, wherein the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier, and the steps include:

[0095] A two-level ensemble learning framework is constructed, wherein the first level of the two-level ensemble learning framework includes a random forest classifier and an extreme gradient boosting classifier, wherein the random forest classifier uses a parameter optimization method based on an outsourcing error rate to determine the number of trees and depth parameters, and the extreme gradient boosting classifier uses a Bayesian optimization algorithm to determine the learning rate and subsampling ratio parameters; based on the prediction outputs of the two classifiers, a weighted voting fusion module is constructed in the second level of the two-level ensemble learning framework, wherein the weighted voting fusion module calculates the prediction result weights of the two classifiers by minimizing the logarithmic loss function, combines the prediction results of the random forest classifier and the extreme gradient boosting classifier according to the optimal weights, and introduces a regularization term in the logarithmic loss function to balance the weight distribution of the two classifiers;

[0096] A time-aware cross-validation strategy is used to train the visitor identification model. The data set is divided into multiple continuous subsets in chronological order. A feature perturbation enhancement mechanism and a sample resampling mechanism are introduced during the training process. The feature perturbation enhancement mechanism adds random noise to numerical features, and the sample resampling mechanism balances the sample distribution by generating synthetic samples.

[0097] An online update mechanism is designed to realize the incremental learning of the visitor identity recognition model. A sample buffer pool based on a sliding time window is maintained. When the number of new samples reaches the preset condition, the incremental learning process is triggered. A progressive update strategy is used to update the parameters of some weak learners. The resampling parameters are dynamically adjusted based on the data distribution of the sample buffer pool, so that the visitor identity recognition model can adapt to the dynamic changes of visitor behavior patterns.

[0098] Exemplarily, in the design of the two-level integrated learning framework, the first level includes two base classifiers, random forest and extreme gradient boosting. The random forest classifier uses a parameter optimization method based on the outsourcing error rate to determine the optimal parameter combination through grid search. In practical applications, through multiple rounds of testing, it was found that when the number of trees is set to 100-150 and the maximum depth of a single tree is set to 8-12 layers, the model performance is optimal. The extreme gradient boosting classifier uses the Bayesian optimization algorithm to find the optimal parameters. After 1000 rounds of iterative optimization, the parameter range of the learning rate is 0.05-0.15 and the feature sampling ratio is 0.6-0.8.

[0099] The second-level weighted voting fusion module guides the weight distribution through the logarithmic loss function. The module input is the predicted probability value of each sample by the two base classifiers, and the output is the final prediction result after fusion. The L2 regularization term is introduced to control the weight distribution to avoid the situation where the weight of a classifier is too large. In practice, the weight ratio of the random forest classifier and the extreme gradient boosting classifier is usually dynamically adjusted between 0.4:0.6 and 0.6:0.4.

[0100] The time-series-aware cross-validation strategy divides 90 consecutive days of data into 6 subsets, each containing 15 days of data. The feature perturbation enhancement mechanism adds Gaussian noise with a mean of 0 and a standard deviation of 0.01 to numerical features to improve the model's robustness to feature perturbations. The sample resampling mechanism uses the SMOTE algorithm to generate synthetic samples, making the ratio of the number of samples in each category close to 1:1, alleviating the problem of uneven sample distribution.

[0101] The online update mechanism maintains a sample buffer pool of 30 days in size. When the number of new samples exceeds 5,000 or the time span exceeds 24 hours, incremental learning is triggered. The progressive update strategy randomly selects 20% of weak learners for parameter update each time. During the update, 90% of the weight is assigned to the new samples, and 10% of the weight is reserved for historical samples to achieve a smooth transition of the model. The sample buffer pool calculates the category weight based on the data distribution of the last 7 days and dynamically adjusts the resampling parameters.

[0102] In the specific implementation scenario, the private traffic visitor identification system of an e-commerce platform processes about 100,000 visitor data every day. In the first level of the two-level integrated learning framework, the random forest classifier contains 120 decision trees with a maximum tree depth of 10 layers, and each tree randomly samples 80% of the features; the extreme gradient boosting classifier sets the learning rate to 0.08, the sampling ratio to 0.7, and the maximum number of iterations to 500. The initial weight ratio of the two classifiers in the second-level fusion module is 0.5:0.5, which is dynamically adjusted during the training process.

[0103] The feature perturbation enhancement mechanism adds random noise to numerical features such as visit duration and click count, with the noise intensity between 1% and 3% of the original value. The sample resampling mechanism generates synthetic samples for low-frequency visitor categories, so that the ratio of the number of samples in each category is controlled between 0.8 and 1.2. The sample buffer pool updates about 3,000 to 5,000 new samples every day, and the incremental learning process is triggered when the cumulative number of samples reaches 5,000.

[0104] The two-level integrated learning framework of the present invention fully integrates the advantages of different classifiers and improves the generalization ability of the model; the time-series-aware cross-validation strategy ensures the time-series consistency of model evaluation; the feature perturbation and sample resampling mechanism enhance the robustness of the model; the online update mechanism realizes the continuous optimization of the model, enabling it to adapt to the changes in visitor behavior patterns in a timely manner. The overall solution shows significant technical advantages in the visitor identification scenario of private domain traffic, providing reliable technical support for precision marketing and personalized services.

[0105] In an optional embodiment,

[0106] Based on the result of the visitor identification model, a multi-level label system is constructed, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption capacity labels, interest preference labels, social relationship labels, and time period labels. The steps include:

[0107] Collect the visitor's device information, IP address and access time, extract the device hardware parameters to form a device feature vector, parse the geographic location coordinates based on the IP address to form a location feature vector, divide the access time into time periods to form a time feature vector, and combine the device feature vector, location feature vector and time feature vector to construct a basic attribute label;

[0108] Applying a bidirectional long short-term memory network to the visitor's page browsing sequence to extract the time series pattern of the behavior trajectory data, calculating the behavior indicators of page dwell time, jump frequency and repeat visit rate, and constructing a behavior feature label based on the time series pattern and behavior indicators, wherein the behavior feature label includes the visitor's browsing type and target characteristics;

[0109] Based on the time feature vector in the basic attribute tag and the browsing type in the behavior feature tag, combined with the transaction amount and purchase frequency of the visitor, a clustering algorithm is used to divide the consumption capacity groups, construct a price sensitivity index, and form a consumption capacity tag;

[0110] Use convolutional neural networks to process visitors' content interaction data to obtain spatial features, extract the distribution of stay time in different content areas, calculate content preference vectors based on click, collection and sharing behaviors, and construct interest preference tags;

[0111] Based on the attention mechanism, the behavioral feature labels are processed to extract the interaction intensity between visitors, build a visitor social graph and calculate the centrality and influence characteristics, and apply the community discovery algorithm to identify the connected visitor groups to form social relationship labels;

[0112] Perform Fourier transform on the access time series to extract periodic features, calculate activity indicators based on the sliding time window, and construct time period labels by combining the time feature vector in the basic attribute label and the target feature in the behavior feature label.

[0113] Exemplarily, the basic attribute label construction process first collects the visitor's device information, including hardware parameters such as operating system version, browser type, screen resolution, etc., and converts these discrete features into device feature vectors through unique hot encoding. The IP address is converted into longitude and latitude coordinates through the geolocation resolution service, and the location feature vector is formed by combining regional information such as city and province. The access time is divided into four time periods of 24 hours: early morning (0-6 o'clock), morning (6-12 o'clock), afternoon (12-18 o'clock), and evening (18-24 o'clock). The access frequency of each time period is counted to form a time feature vector. The three types of feature vectors are spliced ​​to form a complete basic attribute label.

[0114] The extraction of behavioral feature labels uses a bidirectional long short-term memory network. The input dimension is the page browsing sequence of visitors in the past 30 days. The number of hidden units is set to 256. The temporal dependency of the behavior trajectory is captured in the forward and backward directions. The calculation of behavioral indicators includes: the average length of stay on the page, such as staying on the product details page for more than 3 minutes is considered as deep browsing; page jump frequency, counting the number of page switches within 30 seconds; repeat visit rate, calculating the number of repeated visits to the same page within 24 hours. Based on these indicators, visitor behavior is divided into browsing, research, comparison and other types, and the characteristics of their visit purpose are extracted.

[0115] The consumption capacity label is constructed based on the visitor's consumption data for the past 90 days, combined with time characteristics and behavior types. Visitors are divided into three groups: high consumption, medium consumption, and low consumption through the K-means clustering algorithm. The clustering characteristics include: monthly average consumption amount, single maximum consumption amount, purchase frequency, etc. The price sensitivity index is calculated by comparing the price range of the products browsed by visitors and the price range of the products actually purchased, forming three levels of high sensitivity, medium sensitivity, and low sensitivity.

[0116] The interest preference tag uses a five-layer convolutional neural network to process visitor content interaction data. The convolution kernel size is 3x3, and the number of channels is 32, 64, 128, 256, and 512. The proportion of visitors' stay time in different content areas (such as digital, clothing, and beauty products) is counted, and the content preference vector is constructed by combining interactive indicators such as click-through rate, collection rate, and sharing rate. Each area is divided into three levels of strong interest, medium interest, and weak interest according to the intensity of preference.

[0117] The social relationship tag processes the interaction data between visitors based on the multi-head attention mechanism. The number of attention heads is set to 8, and the dimension of each head is 64. The social graph is constructed by counting the comments, replies, likes and forwarding behaviors between visitors. The PageRank algorithm (a method for calculating the importance of web pages based on link relationships) is used to calculate the influence score of visitors, and the Louvain community discovery algorithm (a network community discovery method based on modularity optimization) is used to identify closely connected visitor groups. The social activity, social influence and group belonging indicators are calculated for each visitor.

[0118] The time period tag extracts the periodic pattern of visitor behavior by performing a fast Fourier transform on the access time series. A sliding window with a window size of 7 days is used to calculate the changes in visitor activity at different time scales (day, week, month). Combining the time feature vector and target features, visitors are divided into types such as weekday active, weekend active, and holiday active.

[0119] The present invention achieves a comprehensive characterization of visitor characteristics through the construction of a multi-level label system; the feature extraction method based on deep learning improves the expressiveness of labels; the multi-dimensional label division meets the personalized needs of precision marketing; and the dynamic update mechanism ensures the timeliness of the label system. This solution provides systematic label support for the refined operation of private domain traffic, and can effectively guide the formulation of personalized recommendations and precision marketing strategies.

[0120] In an optional embodiment,

[0121] The steps of extracting features from the visitor's historical behavior trajectory data and historical content interaction data by using a transfer learning algorithm, classifying the extracted features based on the multi-level label system to generate user portrait features, and dynamically updating the user portrait features by using an incremental learning method to construct a multi-dimensional user portrait model include:

[0122] A transfer learning algorithm is used to extract features of visitors' historical behavior trajectory data and historical content interaction data. The historical behavior trajectory data includes click sequence, browsing path and stay time data, and the historical content interaction data includes search terms, comment text and interaction record data. The behavior features and content features are processed separately through a two-stream feature extraction network, and the obtained behavior features and content features are fused through an attention mechanism to obtain a unified feature representation;

[0123] Building a feature mapping network based on a multi-level tag system, decomposing the unified feature representation according to the tag level, building a mapping relationship matrix between features and tags, calculating the weight distribution of features in each tag dimension, and generating structured user portrait features;

[0124] Collect new behavior trajectory data and content interaction data of visitors in real time, design feature importance evaluation indicators, and dynamically update the user portrait features through incremental learning;

[0125] Based on the dynamically updated user portrait features and the feature importance evaluation index, the weight coefficient of the feature mapping network is updated to construct a multi-dimensional user portrait model.

[0126] Exemplarily, the transfer learning algorithm first extracts features from historical behavior trajectory data. For click sequences, a pre-trained Transformer model is used with an input dimension of 128, 8 attention heads, and 6 encoder layers to extract visitor click pattern features. Browsing path data is processed by a graph neural network, where nodes represent visited pages, edges represent jump relationships between pages, and hidden_size is set to 256. Dwell time data is processed by a temporal convolutional network with a convolution kernel size of 3, a step size of 1, and 64 channels.

[0127] The feature extraction of historical content interaction data uses the BERT model (Transformer-based bidirectional encoder representation model) to process search terms and comment texts, with the maximum sequence length set to 512 and the hidden layer dimension set to 768. Interaction record data includes behaviors such as likes, favorites, and shares, which are processed by a multi-layer perceptron, with hidden layer dimensions of 512, 256, and 128, respectively. The two-stream feature extraction network encodes the behavior data and content data respectively, and the output dimension is 512. The attention mechanism uses scaled dot product attention, with the temperature coefficient parameter set to 0.1, to fuse the two types of features into a unified feature representation of 1024 dimensions.

[0128] The feature mapping network is built based on a six-layer label system. The basic attribute label mapping layer contains 256 neurons, the behavioral feature label mapping layer contains 384 neurons, the consumption capacity label mapping layer contains 192 neurons, the interest preference label mapping layer contains 512 neurons, the social relationship label mapping layer contains 320 neurons, and the time period label mapping layer contains 128 neurons. The mapping relationship matrix is ​​optimized by back propagation, and the initial weights are initialized using Xavier. The weight distribution calculation takes into account the correlation and discrimination between the features and labels, and uses the softmax function for normalization.

[0129] The feature importance evaluation adopts a multi-index comprehensive scoring mechanism. The timeliness index calculates the attenuation coefficient according to the data generation time. The weight of data within 7 days is 1.0, the weight of data between 7 and 30 days is 0.8, and the weight of data between 30 and 90 days is 0.5. The stability index is determined by calculating the variance of the eigenvalue in a continuous time window. The smaller the variance, the higher the weight. The discrimination index is calculated based on the entropy of the eigenvalue. The larger the entropy value, the higher the discrimination.

[0130] The real-time data processing module collects new data every 5 minutes. Behavior trajectory data includes the sequence of recently visited pages, duration of stay, and behavioral events. Content interaction data includes search keywords, comment content, and interactive behavior. Incremental learning uses a sliding window mechanism with a window size of 24 hours and updates model parameters every hour. When updating, the new and old features are weighted averaged, and the weight ratio is dynamically adjusted according to the feature importance.

[0131] In a specific application scenario, an e-commerce platform processes about 500,000 pieces of user behavior data every day. When processing behavior data, the dual-stream feature extraction network generates a 512-dimensional behavior feature vector for the typical shopping path "homepage-category page-product list-product details-order". In content data processing, keywords are extracted from user search terms and comment texts and word frequencies are calculated to generate a 512-dimensional content feature vector. The 1024-dimensional feature vector obtained after attention fusion can fully express the user's behavior pattern and content preference.

[0132] The feature mapping network maps the unified feature representation to different label dimensions. For example, the weight distribution of a user's features in the consumption capacity dimension is: high consumption 0.2, medium consumption 0.7, low consumption 0.1; the weight distribution in the interest preference dimension is: digital 0.4, clothing 0.3, beauty 0.3. The feature importance evaluation shows that the recent search behavior and purchase behavior features are more important, with weights between 0.8-1.0; the historical browsing behavior features are less important, with weights between 0.3-0.5.

[0133] The present invention effectively utilizes existing knowledge through transfer learning and improves the efficiency of feature extraction; the dual-stream feature extraction network realizes the deep integration of behavior data and content data; the multi-level label system ensures the comprehensiveness of user portraits; the feature importance evaluation and incremental learning mechanism ensure the real-time and accuracy of the model. This solution can quickly build accurate user portraits and provide strong support for personalized recommendations and precision marketing.

[0134] In an optional embodiment,

[0135] According to the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service node, a routing allocation strategy is used to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service node is monitored. When load imbalance is detected, the steps of dynamically adjusting the routing allocation strategy include:

[0136] A dimensionality reduction algorithm is used to reduce the dimension of user portrait features. Based on the reduced dimensionality user portrait features, a clustering algorithm is used to divide users into multiple clusters, and the mean vector of user portrait features in each cluster is calculated as the representative feature of this type of user;

[0137] Obtain user profile features of new visitors, calculate the matching degree between the user profile features and the representative features of each cluster, and introduce a load balancing factor, wherein the load balancing factor is calculated based on the current load of each service node and the maximum load allowed by the system;

[0138] Constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to a corresponding service node based on the routing decision score;

[0139] The real-time load status of the service node is monitored, and the load standard deviation is calculated. When the load standard deviation exceeds a preset threshold, the routing allocation strategy is dynamically updated by adjusting the weight coefficient of the matching degree and the load balancing factor.

[0140] Exemplarily, the user portrait feature dimensionality reduction process uses the principal component analysis method to reduce the original 1024-dimensional features to 128 dimensions, retaining the features with a principal component contribution rate of 95%. The reduced-dimensional features are grouped using the K-means clustering algorithm, the number of clusters is set to 20, and the silhouette coefficient is used to evaluate the clustering effect. The mean vector of the user portrait feature is calculated for each cluster as the representative feature of this type of user to form a feature center library.

[0141] The cosine similarity metric is used to calculate the matching degree between the new visitor's profile features and the cluster representative features. In the specific implementation, the 128-dimensional feature vector of the new visitor is first normalized, and then the similarity is calculated with the representative features of the 20 clusters respectively. For example, the matching degree of a visitor with the electronic product preference cluster is 0.85, the matching degree with the clothing preference cluster is 0.45, and the matching degree with the beauty preference cluster is 0.35.

[0142] The calculation of the load balancing factor takes into account multiple indicators of the service node. The CPU usage rate is set at 80% as the threshold, and the current usage rate is divided by the threshold to get the CPU load ratio; the memory usage rate is set at 70% as the threshold, and the current usage rate is divided by the threshold to get the memory load ratio; the network bandwidth usage rate is set at 60% as the threshold, and the current usage rate is divided by the threshold to get the bandwidth load ratio. The weighted average of the three indicators is used as the comprehensive load factor of the node. For example, if a node currently has a CPU usage rate of 65%, a memory usage rate of 50%, and a bandwidth usage rate of 45%, the calculated load factor is 0.75.

[0143] The calculation of the routing decision score takes into account the matching degree and the load balancing factor. In the initial configuration, the matching degree weight is 0.6 and the load balancing factor weight is 0.4. The matching degree and the load balancing factor are multiplied by the corresponding weights and added together to obtain the final routing decision score. The system selects the service node with the highest decision score to process the visitor request. For example, for a certain visitor, the matching degree with service node A is 0.85, the node load factor is 0.75, and the calculated routing decision score is 0.79; the matching degree with service node B is 0.75, the node load factor is 0.55, and the calculated routing decision score is 0.67. Finally, the request is assigned to service node A.

[0144] The load monitoring module collects service node status data every 5 seconds. The standard deviation of the comprehensive load factor of all nodes is calculated, and the standard deviation threshold is set to 0.15. When it is detected that the standard deviation exceeds the threshold, the dynamic adjustment mechanism is started. The adjustment process is achieved by changing the weight ratio of the matching degree and the load balancing factor. The greater the load difference, the higher the weight of the load balancing factor. For example, when the load standard deviation reaches 0.2, the matching degree weight is adjusted to 0.4 and the load balancing factor weight is adjusted to 0.6.

[0145] The present invention achieves efficient matching of user characteristics through dimensionality reduction and clustering; the introduction of load balancing factors ensures the rational allocation of system resources; and the dynamic adjustment mechanism ensures the stability of service quality. This solution can achieve optimal utilization of system resources while ensuring user experience, significantly improving service efficiency and system reliability.

[0146] In an optional embodiment,

[0147] The steps of constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to the corresponding service node based on the routing decision score include:

[0148] Obtain a user portrait feature vector and a representative feature vector, and calculate a feature dimension weight, wherein the feature dimension weight is obtained by an information gain method based on the difference between the user category information entropy and the conditional entropy; perform a weighted calculation on the user portrait feature vector and the representative feature vector based on the feature dimension weight to obtain a weighted cosine matching degree;

[0149] Obtaining the current load level of each service node, and normalizing the current load level using a nonlinear function to obtain a load balancing factor;

[0150] Constructing a load balancing loss function including a load variance term and a feature matching deviation term, and calculating a time-varying fusion weight using a gradient descent method based on the load balancing loss function;

[0151] The weighted cosine matching degree and the load balancing factor are weightedly combined by the time-varying fusion weight to obtain a routing decision score;

[0152] The routing decision score is converted into a routing probability distribution by using a soft maximization method, and a selection result of a service node is determined based on the routing probability distribution.

[0153] Exemplarily, the feature dimension weight calculation adopts the information gain method. First, the overall distribution of user categories is statistically obtained to obtain information entropy, and then the category distribution under the condition of each feature dimension is calculated to obtain conditional entropy. The importance of each dimension is determined by the difference between information entropy and conditional entropy. For the 128-dimensional user portrait feature, the calculated dimension weight distribution shows that: the feature dimension related to user consumption capacity has a higher weight, averaging 0.15; the feature dimension related to interest preference has the second highest weight, averaging 0.12; the feature dimension related to basic attributes has a lower weight, averaging 0.08.

[0154] In the calculation process of weighted cosine matching, the user portrait feature vector and the dimension value of the representative feature vector are first multiplied by the corresponding dimension weight. Taking an e-commerce user as an example, the value of the consumption capacity-related dimension in its 128-dimensional feature vector is 0.8, which is multiplied by the weight of 0.15 to get 0.12; the value of the interest preference-related dimension is 0.6, which is multiplied by the weight of 0.12 to get 0.072. After weighting all dimensions, the cosine similarity between the weighted feature vectors is calculated to obtain the final matching value.

[0155] The nonlinear normalization of the service node load level is processed by the sigmoid function. For the three indicators of CPU utilization, memory utilization and network bandwidth utilization, first calculate their ratios to the preset thresholds. The current CPU utilization of a node is 75%, and the ratio to the threshold of 80% is 0.94; the memory utilization is 60%, and the ratio to the threshold of 70% is 0.86; the bandwidth utilization is 50%, and the ratio to the threshold of 60% is 0.83. Substitute these ratios into the sigmoid function for nonlinear transformation to obtain the normalized load balancing factor.

[0156] The load balancing loss function consists of two parts: the load variance term reflects the balance of the system load distribution, and the feature matching deviation term indicates the accuracy of matching the routing allocation with the user features. The fusion weight is dynamically adjusted through the online gradient descent method to minimize the overall loss. In the initial configuration, the weight of the load variance term is 0.4, the weight of the feature matching deviation term is 0.6, the learning rate is set to 0.01, and the weight is updated every 100 requests.

[0157] The calculation of the time-varying fusion weight takes into account the current overall load status of the system and the matching accuracy. When the load is relatively balanced, the weight of feature matching is higher, up to 0.7-0.8; when the load is unbalanced, the weight of the load balancing factor will be increased accordingly, up to 0.6-0.7. In the specific implementation, the exponential sliding average method is used to smooth the weight changes to avoid drastic fluctuations.

[0158] After the routing decision scores are obtained through weighted combination, they are converted into routing probability distribution using the softmax function. For example, a visitor request calculates the routing decision scores of three candidate nodes to be 0.85, 0.75, and 0.65, respectively, and the converted routing probabilities are 0.45, 0.35, and 0.20, respectively. The system gives priority to the node with the highest probability, but retains other candidate nodes as alternatives.

[0159] The present invention improves the matching accuracy by introducing feature dimension weights; nonlinear load normalization reflects the node status more accurately; time-varying fusion weights achieve a dynamic balance between load balancing and feature matching; and soft maximization processing provides a flexible node selection mechanism.

[0160] Figure 2 Schematic diagram of the structure of the visitor identification and multi-dimensional user portrait routing distribution system based on private domain traffic according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0161] The first unit is used to collect multi-source data information of visitors from a private domain traffic pool, wherein the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data, and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features, and behavior features, perform feature fusion and cluster analysis on the multi-dimensional feature vector using a multi-branch deep neural network model, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features;

[0162] The second unit is used to construct a multi-level label system based on the result of the visitor identity recognition model, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption ability labels, interest preference labels, social relationship labels, and time period labels; a transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features, and the user portrait features are dynamically updated through incremental learning to construct a multi-dimensional user portrait model;

[0163] The third unit is used to allocate different visitor requests to corresponding service nodes based on the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service nodes using a routing allocation strategy, and monitor the real-time load status of the service nodes at the same time. When load imbalance is detected, the routing allocation strategy is dynamically adjusted.

[0164] According to a third aspect of the embodiments of the present invention,

[0165] An electronic device is provided, comprising:

[0166] processor;

[0167] a memory for storing processor-executable instructions;

[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] A fourth aspect of the embodiments of the present invention is:

[0170] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0171] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for visitor identification and multi-dimensional user portrait routing allocation based on private domain traffic, characterized in that: include: Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features; Based on the results of the visitor identification model, a multi-level label system is constructed, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption capacity labels, interest preference labels, social relationship labels, and time period labels; A transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features. The user portrait features are dynamically updated through incremental learning to build a multi-dimensional user portrait model. According to the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service nodes, a routing allocation strategy is used to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service nodes is monitored. When load imbalance is detected, the routing allocation strategy is dynamically adjusted. According to the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service node, a routing allocation strategy is used to allocate different visitor requests to the corresponding service nodes. At the same time, the real-time load status of the service node is monitored. When load imbalance is detected, the steps of dynamically adjusting the routing allocation strategy include: A dimensionality reduction algorithm is used to reduce the dimension of user portrait features. Based on the reduced dimensionality user portrait features, a clustering algorithm is used to divide users into multiple clusters, and the mean vector of user portrait features in each cluster is calculated as the representative feature of this type of user; Obtain user profile features of new visitors, calculate the matching degree between the user profile features and the representative features of each cluster, and introduce a load balancing factor, wherein the load balancing factor is calculated based on the current load of each service node and the maximum load allowed by the system; Constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to a corresponding service node based on the routing decision score; The real-time load status of the service node is monitored, and the load standard deviation is calculated. When the load standard deviation exceeds a preset threshold, the routing allocation strategy is dynamically updated by adjusting the weight coefficient of the matching degree and the load balancing factor.

2. The method according to claim 1, characterized in that Collect multi-source data information of visitors from the private domain traffic pool, the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features and behavior features, use a multi-branch deep neural network model to perform feature fusion and cluster analysis on the multi-dimensional feature vector, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features, including the following steps: Preprocessing the multi-source data information, wherein the time holes in the behavior trajectory data are filled by an interpolation method based on temporal correlation, the abnormal parameters in the device feature data are detected and processed by the interquartile range method, the numerical features are normalized by minimum and maximum, and the text data are segmented and vectorized; Constructing a multidimensional feature vector based on the preprocessed multi-source data information, the multidimensional feature vector includes time series features, spatial features and behavioral features, wherein the time series features of behavior frequency and duration are extracted from the behavioral trajectory data, the spatial features of activity areas and device migration trajectories are constructed using the device feature data and geographic location data, and the behavioral features of content preferences and interaction patterns are extracted in combination with the social relationship data and content interaction data; The multi-dimensional feature vector is subjected to feature fusion by using a multi-branch deep neural network model, wherein the multi-branch deep neural network model includes a temporal feature branch, a spatial feature branch and a behavioral feature branch, wherein the temporal feature branch uses a bidirectional long short-term memory network to process temporal features, the spatial feature branch uses a convolutional neural network to process spatial features, and the behavioral feature branch uses an attention mechanism to process behavioral features, and the temporal features, spatial features and behavioral features are fused by a weight fusion mechanism; An improved density-based spatial clustering algorithm is used to perform cluster analysis on the fused features. The improved density-based spatial clustering algorithm includes: using a kernel density estimation method to adaptively calculate a neighborhood radius parameter, dynamically adjusting a minimum sampling point threshold based on a time window sliding average method, introducing a timestamp-based neighborhood constraint, and generating a visitor identity fingerprint feature that includes visitor behavior patterns, activity patterns, and social characteristics; A visitor identification model is established based on the visitor identity fingerprint feature, and the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier.

3. The method according to claim 2, characterized in that A visitor identification model is established based on the visitor identity fingerprint feature, wherein the visitor identification model adopts an integrated learning framework to fuse a random forest classifier and an extreme gradient boosting classifier, and the steps include: A two-level ensemble learning framework is constructed, wherein the first level of the two-level ensemble learning framework includes a random forest classifier and an extreme gradient boosting classifier, wherein the random forest classifier uses a parameter optimization method based on an outsourcing error rate to determine the number of trees and depth parameters, and the extreme gradient boosting classifier uses a Bayesian optimization algorithm to determine the learning rate and subsampling ratio parameters; based on the prediction outputs of the two classifiers, a weighted voting fusion module is constructed in the second level of the two-level ensemble learning framework, wherein the weighted voting fusion module calculates the prediction result weights of the two classifiers by minimizing the logarithmic loss function, combines the prediction results of the random forest classifier and the extreme gradient boosting classifier according to the optimal weights, and introduces a regularization term in the logarithmic loss function to balance the weight distribution of the two classifiers; A time-aware cross-validation strategy is used to train the visitor identification model. The data set is divided into multiple continuous subsets in chronological order. A feature perturbation enhancement mechanism and a sample resampling mechanism are introduced during the training process. The feature perturbation enhancement mechanism adds random noise to numerical features, and the sample resampling mechanism balances the sample distribution by generating synthetic samples. An online update mechanism is designed to realize the incremental learning of the visitor identity recognition model. A sample buffer pool based on a sliding time window is maintained. When the number of new samples reaches the preset condition, the incremental learning process is triggered. A progressive update strategy is used to update the parameters of some weak learners. The resampling parameters are dynamically adjusted based on the data distribution of the sample buffer pool, so that the visitor identity recognition model can adapt to the dynamic changes of visitor behavior patterns.

4. The method according to claim 1, characterized in that Based on the result of the visitor identification model, a multi-level label system is constructed, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption capacity labels, interest preference labels, social relationship labels, and time period labels. The steps include: Collect the visitor's device information, IP address and access time, extract the device hardware parameters to form a device feature vector, parse the geographic location coordinates based on the IP address to form a location feature vector, divide the access time into time periods to form a time feature vector, and combine the device feature vector, location feature vector and time feature vector to construct a basic attribute label; Applying a bidirectional long short-term memory network to the visitor's page browsing sequence to extract the time series pattern of the behavior trajectory data, calculating the behavior indicators of page dwell time, jump frequency and repeat visit rate, and constructing a behavior feature label based on the time series pattern and behavior indicators, wherein the behavior feature label includes the visitor's browsing type and target characteristics; Based on the time feature vector in the basic attribute tag and the browsing type in the behavior feature tag, combined with the transaction amount and purchase frequency of the visitor, a clustering algorithm is used to divide the consumption capacity groups, construct a price sensitivity index, and form a consumption capacity tag; Use convolutional neural networks to process visitors' content interaction data to obtain spatial features, extract the distribution of stay time in different content areas, calculate content preference vectors based on click, collection and sharing behaviors, and construct interest preference tags; Based on the attention mechanism, the behavioral feature labels are processed to extract the interaction intensity between visitors, build a visitor social graph and calculate the centrality and influence characteristics, and apply the community discovery algorithm to identify the connected visitor groups to form social relationship labels; Perform Fourier transform on the access time series to extract periodic features, calculate activity indicators based on the sliding time window, and construct time period labels by combining the time feature vector in the basic attribute label and the target feature in the behavior feature label.

5. The method according to claim 1, characterized in that The steps of extracting features from the visitor's historical behavior trajectory data and historical content interaction data by using a transfer learning algorithm, classifying the extracted features based on the multi-level label system to generate user portrait features, and dynamically updating the user portrait features by using an incremental learning method to construct a multi-dimensional user portrait model include: A transfer learning algorithm is used to extract features of visitors' historical behavior trajectory data and historical content interaction data. The historical behavior trajectory data includes click sequence, browsing path and stay time data, and the historical content interaction data includes search terms, comment text and interaction record data. The behavior features and content features are processed separately through a two-stream feature extraction network, and the obtained behavior features and content features are fused through an attention mechanism to obtain a unified feature representation; Building a feature mapping network based on a multi-level tag system, decomposing the unified feature representation according to the tag level, building a mapping relationship matrix between features and tags, calculating the weight distribution of features in each tag dimension, and generating structured user portrait features; Collect new behavior trajectory data and content interaction data of visitors in real time, design feature importance evaluation indicators, and dynamically update the user portrait features through incremental learning; Based on the dynamically updated user portrait features and the feature importance evaluation index, the weight coefficient of the feature mapping network is updated to construct a multi-dimensional user portrait model.

6. The method according to claim 1, characterized in that The steps of constructing a routing allocation strategy according to the matching degree and the load balancing factor, performing weighted calculation on the matching degree and the load balancing factor to obtain a routing decision score, and allocating the request of the new visitor to the corresponding service node based on the routing decision score include: Obtain a user portrait feature vector and a representative feature vector, and calculate a feature dimension weight, wherein the feature dimension weight is obtained by an information gain method based on the difference between the user category information entropy and the conditional entropy; perform a weighted calculation on the user portrait feature vector and the representative feature vector based on the feature dimension weight to obtain a weighted cosine matching degree; Obtaining the current load level of each service node, and normalizing the current load level using a nonlinear function to obtain a load balancing factor; Constructing a load balancing loss function including a load variance term and a feature matching deviation term, and calculating a time-varying fusion weight based on the load balancing loss function using a gradient descent method; The weighted cosine matching degree and the load balancing factor are weightedly combined by the time-varying fusion weight to obtain a routing decision score; The routing decision score is converted into a routing probability distribution by using a soft maximization method, and a selection result of a service node is determined based on the routing probability distribution.

7. A visitor identification and multi-dimensional user portrait routing distribution system based on private domain traffic, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect multi-source data information of visitors from a private domain traffic pool, wherein the multi-source data information includes behavior trajectory data, device feature data, geographic location data, social relationship data, and content interaction data, pre-process the multi-source data information, construct a multi-dimensional feature vector including time series features, spatial features, and behavior features, perform feature fusion and cluster analysis on the multi-dimensional feature vector using a multi-branch deep neural network model, generate visitor identity fingerprint features, and establish a visitor identity recognition model based on the visitor identity fingerprint features; The second unit is used to construct a multi-level label system based on the result of the visitor identity recognition model, wherein the multi-level label system includes basic attribute labels, behavior feature labels, consumption ability labels, interest preference labels, social relationship labels and time period labels; A transfer learning algorithm is used to extract features from the visitor's historical behavior trajectory data and historical content interaction data, and the extracted features are classified based on the multi-level label system to generate user portrait features. The user portrait features are dynamically updated through incremental learning to build a multi-dimensional user portrait model. The third unit is used to allocate different visitor requests to corresponding service nodes based on the user portrait features of the multi-dimensional user portrait model and the real-time load status of the service nodes using a routing allocation strategy, and monitor the real-time load status of the service nodes at the same time. When load imbalance is detected, the routing allocation strategy is dynamically adjusted.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images